add softmax kernel 9: online softmax with 4x unrolled loads

[RahulPatnaik]

Add softmax forward kernel 9

Adds a new kernel variant to dev/cuda/softmax_forward.cu that builds on kernel 8 (online softmax with warp-level reductions).

What changed

Kernel 9 unrolls the coarsening loops by a factor of 4, issuing multiple global memory loads together to increase memory-level parallelism. The online softmax algorithm (fused max + sum in a single pass) and warp shuffle reductions are kept from kernel 8. A precomputed reciprocal replaces per-element division in the output pass.

Results (RTX 4060 Laptop GPU)

Kernel	Time	DRAM Throughput
8	22.82 ms	84.45%
9	20.76 ms	92.82%

~9% wall-clock improvement. The gain comes from better memory bus utilization through overlapped loads, not from reduced memory traffic (both kernels make the same number of global memory passes).

Notes

• Correctness validated against CPU reference at all block sizes (32, 64, 128, 256, 512, 1024)
• Uses the min(C-1, ...) out-of-bounds clamping trick from kernel 7 to keep loads branchless
• No shared memory required (warps are independent, one warp per row)